an estimable model of supermarket behavior: …betancourt/distribution/estimable.pdfan estimable...
TRANSCRIPT
An Estimable Model of Supermarket Behavior: Prices, Distribution Services and Some Effects
of Competition*
Roger R. Betancourt and Margaret Malanoski Economics Department Office of Information and Regulatory Affairs U. of Maryland Office of Management and Budget College Park, MD 20742 Washington, DC Ph: 301 4053479 Fax:301 4053542 Revised January 1999
Abstract In this paper we present and estimate a simple model of supermarket behavior that has several attractive properties: It permits the incorporation of the (distribution) services provided by a supermarket as an output of supermarkets and a determinant of demand for supermarket products; it generates, as a special case, one of its main competitors in the supermarket literature -- the so called full price model of services; and, it can be estimated with a unique data set originally constructed by the Economic Research Service of USDA. The main results of the analysis are three. First, the aggregate demand for a supermarket=s products depends critically on distribution services: at the substantive level, a 1 % increase in these services increase quantity demanded by 0.4%; at the methodological level, the restrictions on the parameter values implied by the model are critical in the evaluation of functional forms for demand. Second, supermarkets exhibit constant marginal costs with respect to the quantity of output or turnover and substantially declining marginal costs with respect to (distribution) services, which implies substantial multiproduct economies of scale. Third, in response to an exogenous increase in competition those supermarkets that have adopted newer formats such as superstores and that employ newer technology such as optical scanners choose prices and (distribution ) services in ways that increase consumer welfare, whereas those that do not have these characteristics choose prices and services in ways that lower consumer welfare. JEL Classification: L81, L1, L4. Key Words: Supermarkets, prices, distribution services, demand; economies of scale.
I. Introduction.
2
Supermarkets are an important component of modern economies. For
instance, according to the U.S. Statistical Abstract (1998), in 1996 the
retail trade accounted for 9.37% of U.S. GDP and supermarket sales
accounted for 12.3 % of sales in the retail trade. At the same time the
supermarket category has undergone major structural changes in the 1980's,
for example according to the U.S. Statistical Abstract (1991) the share of
sales of conventional supermarkets went from 73.1% in 1980 to 42% in 1989.
Newer formats gained the ground lost by conventional supermarkets. These
formats provide either broader and/or deeper assortments (for example
superstores, combination food and drug and hypermarkets), or in the case of
warehouses lower prices in exchange for less assurance of product delivery
in the desired form (large packages rather than small ones). By 1996, the
share of conventional supermarkets had fallen to 22.2%, U.S. Statistical
Abstract (1998).
In the trade literature the importance of the services provided by
supermarkets is frequently mentioned. For instance in a 1987 survey of
customers reported by The Progressive Grocer , prices were only one of 16
items considered important by half of the sample in selecting a food store.
The others were services of various kinds such as availability of a produce
department, unit pricing, convenient location, cleanliness, short waits at
checkout counters, etc. Similar results appear in other years, see for
example the April issue in 1993. By and large this aspect of supermarkets
has been either ignored or tangentially acknowledged in the mainstream
economics literature, including the industrial organization literature.
3
The major aim of this paper is to capture the importance of these
services as an economic variable that determines supermarket behavior both
theoretically and empirically. We will do so by treating them explicitly
as an output of supermarkets and as a determinant of consumers= demand for
supermarket products. In Section II, we present a simple theoretical model
that adapts the one in Betancourt and Gautschi (1993) for empirical
implementation with the data available to us. We show how this model
generates as special cases the standard textbook monopoly model, a model of
retail pricing proposed by Bliss (1988), and the perfectly competitive
model referred to as the full price model of services. The latter was put
forth by Ehrlich and Fisher (1982) to analyze the demand for advertising
and has been applied to supermarkets by Oi (1992) and to retailing in
general by Divankar and Ratchford (1995). Our simple model can be viewed
as a slight generalization of the Bertrand model that implies supermarkets
behave as imperfect competitors.
What makes this project feasible is a unique data set put together by
the Economic Research Service of USDA, see Kaufman and Handy(1989).
Supermarkets often provide over a 100,000 products and it is not easy to
obtain an average price charged by a supermarket; they also offer a variety
of services and it is not easy to obtain information on them. By combining
price surveys with a store survey, this data set allows the construction
for each supermarket of an index of prices and an index of services. The
latter correspond to the ones identified as distribution services in the
retailing literature from which the model is derived. Since this data was
4
also combined with social and economic information at the zip code level,
it becomes possible to estimate an aggregate demand function for each
supermarket that incorporates price and distribution services as endogenous
variables and that must satisfy certain theoretical restrictions implied by
the model. The results of this estimation are reported in Section III.
They show that distribution services are an important determinant of the
demand for supermarket products and that not all functional forms for the
demand functions satisfy the restrictions implied by the theory.
From the simple model one can extract the first-order conditions that
determine the choice of prices and distribution services by a supermarket.
By imposing functional forms on cost functions and on the limited
characterization of competition in the model, it becomes possible to
estimate these first-order conditions given the aggregate demand for the
supermarket products. These functional forms and their implications for
estimation are discussed in Section IV. Of particular interest is allowing
for the possibilities of increasing returns to scale, since this issue is
controversial in the literature. Some authors argue in favor of their
existence, Oi(1992) and Ofer (1973); other authors argue in favor of
constant returns to scale, Ingene (1984) and Cotterill(1991). We find that
both groups are right. That is we find constant returns to scale with
respect to turnover or the standard output measure but increasing returns
to scale with respect to distribution services or the output measure we are
emphasizing here.
5
The last section of the paper presents the results, discusses various
econometric issues and reports a simulation: namely, the effect of a
change in the exogenously given level of competition for each observation
(supermarket) using the estimated parameter values. The simulation results
are interesting for two reasons. First, they confirm the main criticisms
of the traditional linear regressions of prices against measures of market
power in grocery retailing made by Anderson (1990). He argues that a
change in competition may not have the same effect on prices in all markets
and we find that prices sometime increase and sometimes decrease; he also
argues that it is necessary to control for differences in quality (what we
call distribution services) and we find that these services also sometimes
increase and sometimes decrease. Second, we find that our sample splits
along interesting characteristics in their response to a change in
competition. For instance supermarkets where consumer welfare increases
have a higher proportion of superstores and use of optical scanners than
those where welfare decreases, which suggests the structural changes we
have observed in the 1980's have been beneficial to consumers.
II. The Model and Its Main Special Cases.
The essence of the model is that retailers choose prices and
distribution services simultaneously given the demand function for
supermarket products and the level of competition in the market. Formally,
we write the constrained maximand for a supermarket as
L = pQ - C(v, Q, D) - wQ +µN[E=- E(p, p=, D, z0 )]
6
where p is a store=s average retail price; Q is the level of output, which
is determined by the aggregate demand faced by the store. N is the number
of transactions. We will assume that all repeated purchases are the same
and that all consumers of any one supermarket are identical; hence, Q= qN,
where q is the demand function per transaction of the representative
consumer, and N will be normalized at unity. C is a neoclassical cost
function describing the costs of supermarket activities as a function of
input prices (v) and the two outputs of this retailing activity, explicit
products and services (Q) and the implicit levels of distribution services
(D). w is the average price of the explicit products and services
purchased from suppliers. E is the expenditure function of the
representative consumer, which depends on the store=s average retail price
(p), the distribution services of the supermarket (D), other prices ( p=)
faced by the consumer and the optimal level of the consumption activities,
z0.1 E= is the lowest cost to this 'representative consumer' of attaining
her maximum level of utility at an alternative establishment. µ is a
Lagrange multiplier.
Since this constrained maximand is unfamiliar to many economists, it
is instructive to discuss what it captures explicitly. The first three
terms are the standard definition of profits for a retail establishment, in
which it is useful to separate sales (pQ) and two dimensions of costs:
those due to the distribution activities of the retailer ( C(v, Q, D)) and
those due to the production activities of wholesalers or suppliers (wQ).
The fourth term represents a constraint that captures the effects of
7
competition on supermarket behavior in the following sense. If there is a
lowering of the competitive standard that a store faces, for example by a 1
$ increase in the lowest cost per transaction to a 'representative
consumer' of attaining her maximum level of utility at an alternative
establishment, the Lagrange multiplier, µ, measures the marginal
contribution to profits of such an experiment. µ ranges between zero and
unity.2 This constraint captures Bliss (1988) concept of retail
competition as offering the consumer better value for her money than at the
next best alternative store in terms of an expenditure function that allows
for the existence of distribution services and makes explicit the
assumption of identical transactions across consumers and repeat
purchases.3
Optimal choices of prices and services by a supermarket must satisfy
Price: (p - CQ - w)(MQ/Mp) + Q(1-
µ) = 0 (1)
Distribution Services: (p - CQ -w)(MQ/MD) +µr - CD = 0
(2)
Constraint: E= - E(p, p=, D, z0 )
= 0 (3).
It is insightful to proceed by considering several special cases that
this model generates, which can be seen through these conditions. Suppose
we assume first that µ= 0. The constraint becomes irrelevant and we have
a generalization of the standard monopoly case in the literature. (1) and
8
(2) simply imply that the retailer chooses prices and distribution services
such that marginal revenues equal marginal costs in both cases.
Suppose we assume instead that µ= 1. Condition (1) implies p = CQ +
w, or the average retail price equals the marginal cost of retailing an
additional unit of output plus the cost of purchasing this unit from
suppliers; condition (2) implies that r = CD, or the shadow price of an
additional unit of distribution services to the representative consumer
equals the marginal cost of providing this additional unit to the consumer.
Thus, we have a generalization of standard competitive behavior in a model
that incorporates distribution services as a variable. Notice that in this
model the conjectural variation is zero. That is the supermarket takes its
rivals prices and distribution services as given.4 Hence, this model has
all the advantages and disadvantages of the Bertrand model as discussed,
for example, in Tirole (1988, Ch. 5). In particular, the slightest
departure from competitive behavior results in the supermarket losing all
of its sales to the representative consumer. This helps interpret the more
general model. Namely, when 0 < µ < 1, the supermarket loses a fraction
of its sales to the representative consumer when it fails to meet the
competitive standard represented by the constraint.
By making an additional assumption in this special case when µ = 1,
it is possible to generate the full price model of services. Ehrlich and
Fisher (1982) argued that perfect competition implies that the full price
paid by the consumer must be the same at every store. Retailers can
compete by offering whatever combination of services and prices they want
9
as long as they meet this constraint. The full price in our model is the
sum of the retail price and the shadow price of distribution services, p +
r. The implication of Ehrlich and Fisher=s argument is that this sum is
equal to a constant, let us say K, which must be the same for every store.
From equation (1) and (2) when µ= 1, this also implies that p + r = K = CQ
+w + CD . The second equality brings out a little known feature of the
full price model. Namely if there are decreasing marginal costs with
respect to distribution services (information in Ehrlich and Fisher=s
case), for example, the model breaks down. That is, given K, increases in
p and decreases in r imply that CQ decreases (if marginal costs are
increasing with respect to output or turnover, Q) and CD decreases (if
marginal costs are decreasing with respect to distribution services, D) and
the second equality can not be met.
III. Specification and Estimation of the Aggregate Demand Function for a
Supermarket.
Implementing the previous model empirically requires estimation of
the aggregate demand function for a supermarket. In this section we
specify and estimate this aggregate demand function, taking into account
the restrictions implied by the previous model. Equation (4) below
specifies the aggregate demand function for a supermarket,
Q = g(p, D, Y, X)u3 . (4).
Q is the level of demand, which is measured as the annual level of sales of
the store deflated by the store=s average price. u3 is a disturbance term,
which is assumed to be lognormally distributed so that lnu3 is normally
10
distributed with E(lnu3 ) = 0. We follow the applied literature by
specifying g in ACobb-Douglas@ form so that it generates elasticities as
parameters and a 'double log' type of demand function5 , i.e.,
lnQ = α + δ1 lnp + δ2 lnD +δ3 lnY +δ4 lnX +ln u3.
(5)
One modification of (5) is useful in an empirical setting because it
allows us to take advantage of one of the strengths of our data, which
permit the representative consumer faced by a supermarket to differ across
supermarkets. Namely, the elasticities affecting the endogenous variables,
p and D, need not be assumed constant. Instead, they will be allowed to
vary with characteristics of the households in the zip code area in which a
supermarket is located, that is
δ1 = δ10 +δ11 X1 < -1 (6)
0 # δ2 = δ20 +δ21 X2 < 1. (7)
Economic theory leads us to expect a negative price response and a
positive distribution services response. Thus, we would expect δ10 < 0 and
δ20 > 0 . The signs of δ11 and δ21, however, would be determined by whether
we expect the household characteristic to increase or decrease the
sensitivity of demand to retail prices or distribution services,
respectively. In the empirical analysis we will assume X1 to be the
percentage of households without cars in the zip code area where a
supermarket is located; hence, we would expect it to decrease the absolute
value of the price elasticity of demand, so that we would expect δ11 >0
because the lesser the access to a car of the representative consumer the
11
less price elastic is demand. This allows us to investigate empirically
one of the mechanisms by which supermarkets may charge different prices to
poor households. We will assume X2 to be the socioeconomic status of the
zip code area where the supermarket is located.6 The inequalities in (6)
and (7) have to be satisfied or second-order conditions in the model of
Section 2 would be violated. This provides a test of the empirical
judgements made, including the selection of X1 and X2 . Y was measured as
the median income in the zip code area where the supermarket was located
and δ3 is, of course, expected to be positive.
X was specified as a vector of three exogenous variables: a dummy
variable indicating whether or not the store was in a shopping center (X41
),7 the population of the zip code area where the store was located (X42 )
and the selling area of the supermarket measured in squared feet of store
space devoted to selling (X43 ). Selling area captures the consequences of
an estimate of the level of demand made by those who designed the
supermarket and unobserved by the researcher and it should be positively
related to the level of demand. Population should be positively related to
the level of demand facing the supermarket if everything else in the zip
code area were equal. Finally, the dummy controls for one of the things
that may not be equal: namely, the pattern of store traffic may be
different when a store is located in a shopping center so that it may have
a higher or lower level of demand given population and selling area.
The Economic Research Service (ERS) of USDA has developed a unique
data base which is the basis for our analysis. It consists of three price
12
surveys (waves 1,2,3) taken six weeks apart in 1982; a separate survey of
store characteristics undertaken over the same period; demographic and
socioeconomic information for zip code areas purchased by ERS from Claritas
Corporation; and SMSA data. We gathered data on the number of food stores
in each SMSA. A detailed description of the data is provided in an
Appendix available upon request.
The data is essentially cross-section data with the unusual feature
that for one variable there are three observations or drawings. That is,
each of the three price surveys generated a store price index, p. Not all
the stores were the same in each price survey, so we worked with a sample
of 430 observations that were included in each wave and checked our results
against the wider samples. The annual sales of each store and the price
index were used to generate Q for each wave. The store characteristics
survey generated the data to construct an index of distribution services
for each store, D, based on the response to twenty questions about whether
or not the store provided a particular service. In addition, for the
estimation of demand the variables X41 and X43 were taken from the store
survey while X1, X2, Y, and X42 were taken from the Claritas data set.
Descriptive statistics on these variables as well as on those introduced in
the next section are presented in Table 1.
>>>>>Table 1 GOES HERE>>>>>>
Table 2 presents the results of estimating (5) with the modifications
implied by (6) and (7) for each of the three waves. The equation to be
estimated is linear in the parameters but nonlinear in the endogenous
13
variables (p and D); hence, it was estimated by nonlinear two stage least
squares.8 The results are not very sensitive to the price survey used in
the analysis. In all three cases, inequalities (6) and (7) are satisfied
without violations at any data point. The coefficients of price,
distribution services, the interaction between price and percentage of
households without a car, median income and selling area have the expected
sign and are statistically significant at well beyond the 1% level in all
three price surveys. The coefficients of population and the shopping
center dummy are not statistically significant at the 2.5% and at the 5%
level, respectively, in any of the three waves. Finally, the interaction
between distribution services and socioeconomic status is statistically
significant at the 5% level in waves 1 and 2 and at the 10% level in wave
3.
>>>>>Table 2 GOES HERE>>>>
Our results show that it is feasible empirically to incorporate
distribution services into the estimation of demand functions for
supermaket products in a theoretically consistent fashion. Substantively, a
1% increase in store price decreases quantity demanded by about 2.0 to 2.2
% on average and a 1% increase in store distribution services increases
quantity demanded by about 0.38 to 0.42% on average. In addition, we find
that one of the mechanisms through which the poor face higher prices is
that their price elasticity of demand is smaller (in absolute value) when
they reside in households without a car. Not surprisingly, establishments
built to sell more do so, but not proportionately ( the hypothesis that the
14
coefficient is unity is rejected at the 1% level); on the other hand,
households with a higher level of median income buy proportionally more at
the supermarket. Higher socioeconomic status in an area reduces the
magnitude of the increase in demand from increasing distribution services,
perhaps because households in these areas take for granted a high level of
distribution services.
In the course of our analysis we performed a number of experiments,
some of which are worth reporting. We set the interaction terms to zero
(δ11 = δ21 =0). The estimated price elasticity was negative and
statistically significant at the 5% level, but less than unity in absolute
value which violates second-order conditions. We also estimated our
specification with SMSA dummies added. The price elasticity estimate was
negative but the null hypothesis that it was greater than or equal to minus
unity could not be rejected at any reasonable level of significance. We
estimated the specification with (not shown) and without an intercept (
Table 2). The intercept estimate is quite large and statistically
insignificant (t < 1), and the null hypothesis that the coefficient of
distribution services is greater than or equal to unity can not be rejected
at any reasonable level of significance for two of the waves and at the 1%
level for all three waves. All three alternatives were discarded because
of their inconsistency with the theoretical restrictions emanating from the
simple model of Section II.
IV. Specification of the First- Order Conditions.
15
Because the model presented in Section II is a perfect information
one, we introduce error terms in equations (1) and (2), u1 and u2
respectively, by assuming that the source of the errors is in the
application of the decision rules for optimization by each agent
(supermarket), not in their perceptions of the objective function. These
considerations lead to the following set of equations for estimation.9
Price: (p - CQ -w) (MQ/ Mp)
+Q (1-µ) = u1 (8)
Distribution Services: (p - CQ - w)(MQ/MD) + µr - CD
= u2 (9).
We will assume E(ui ) = 0, and E(u1u2 ) … 0. That is, the errors in the two
equations capturing the decisions of a supermarket with respect to prices
and distribution services are likely to be correlated and will be allowed
to be in the estimation.10
In order to estimate (8) and (9), we need to specify the marginal
costs with respect to Q and D and the (exogenous) level of competition
faced by a supermarket (µ). Since Q depends on the endogenous variables,
p and D, it will be treated as an endogenous variable in the estimation of
(8) and (9). While equations (8) and (9) have the same form for every
supermarket, their values vary across supermarkets because aggregate
demand, the slopes of the demand functions, marginal costs and the level of
competition can vary across supermarkets. The disturbance terms will be
assumed to be independent and identically distributed across supermarkets.
16
In order to proceed, we need to specify functional forms for the
marginal cost functions for explicit products, CQ, and distribution
services, CD. Our choice was guided by several considerations. First,
the strength of our data is in the measurement of the two aggregate outputs
of supermarkets, Q and D. Second, whether marginal costs are increasing or
decreasing in these outputs is an important consideration in the full price
model, as shown in Section II, but the logic extends to the more general
model. Finally, as noted in the introduction, the issue of whether or not
there are returns to scale in retailing, including supermarkets, has
attracted considerable attention in the literature. Of the multiproduct
cost functions suggested in the standard reference in the industrial
organization literature for multiproduct cost functions, Baumol, Panzar and
Willig (1982, Ch. 15), the one that seemed most suitable in light of these
considerations is a slight generalization of the quadratic cost function
attributed to Braunstein by Baumol and Braunstein (1977).
It generates the following marginal cost functions,
CQ = (exp(θS))[α1 + α11 β Q(β
- 1) + α12 D][v1
π(1) v2 π(2) ]
(10)
CD = (exp(θS))[α2 + α22 λD(λ - 1) + α12 Q][v1
π(1) v2 π(2) ]
(11)
where v1 is occupancy cost, constructed on the basis of SMSA data, and v2
is labor compensation, which was measured in terms of an index of within
SMSA=s variations for each store.11 S is a vector of shift variables that
lead to differences in the levels of costs, for example store type (S1 is a
17
dummy for superstores and S2 is a dummy for traditional supermarkets with
warehouses as the residual category) or the existence of a scanner at the
store (S3). θ is a vector of corresponding coefficients. These four
variables were taken from the store survey.
This specification generalizes the standard quadratic form. That is,
if β = 2 = λ, it collapses to the quadratic. It allows for richer
behavior in terms of multiproduct returns to scale and the shape of
marginal costs than permitted by the standard quadratic.12 For instance,
with these functions marginal costs can increase at either an increasing or
a decreasing rate with distribution services or output as λ or β is
greater or less than one, respectively. Moreover the standard definition
of multiproduct returns to scale as the proportionate increase in costs as
a result of a proportionate increase in outputs, RTS, yields in this case
RTS = 1 + [(β -1)α11 Q + (λ - 1)α22 D]/C.
(12)
Thus, we can easily identify whether returns to scale, if any, are being
generated by a decreasing marginal cost function with respect to
distribution services or with respect to output.
Finally, a data limitation imposes the need to estimate a component
of cost, namely the wholesale price. The latter was assumed to be a
function of whether or not the supermarket belonged to a chain as follows :
w = σ0 + σ1S4, where S4 takes on the value of unity if the store reported
belonging to a chain in the store survey and zero otherwise. σ0 is expected
18
to be positive and σ1 is expected to be negative, i.e., one of the benefits
of belonging to a chain is to secure products at advantageous prices. One
way of checking the reasonableness of this procedure is to check whether or
not the estimated margin, p- CQ - w, for any observation is negative, which
would violate second-order conditions. Fortunately, our data and
estimation procedures allow us to perform this check.
Just as indicated in Section II, the Lagrange multiplier (µ) captures
the level of competition faced by a supermarket as the fraction of sales to
the representative consumer diverted to other stores from a failure to meet
the competitive standard. Hence, one would expect it to vary across the
market areas13 where the supermarkets are located as characteristics of
these market areas differ. Thus, while the value of µ faced by any
supermarket is assumed constant, the value this constant takes on is
allowed to differ across supermarkets. In addition, the value that µ can
take on for any supermarket must lie between zero and unity. These
considerations were incorporated into the empirical analysis by specifying
the following logistic functional form
µ = eγM /(1 + eγM ),
(13)
where M is a vector of variables describing the characteristics of the
market area.14
One characteristic of a market area is market growth, which will be
measured as the rate of growth of food store sales over the previous five
years (M3 ). We would expect that a market area with high market growth,
19
given the same number of stores for example, would have a lower value of
µ than one with low market growth. That is the fraction of sales to the
representative consumer of any one supermarket diverted to other stores as
a result of failing to meet the competitive standard would be less in the
market area with high growth, which implies γ3 < 0 since Mµ/ MMj = γj µ(1 -
µ).
Two other variables were used in the vector M. The number of food
stores per 1000 persons in the SMSA where the supermarket is located, M2 .
One would expect that the higher this number, given market growth for
example, the higher the fraction of sales to the representative consumer of
anyone supermarket in the area that can be diverted to other stores as a
result of a failure to meet the competitive standard, i.e., γ2 > 0.
Finally, the last variable included in this vector was the market share of
the firm that owns the store in the SMSA where the store is located, M1 .15
Statistically, this variable is appealing because, in contrast to the
previous two, it varies across supermarkets within an SMSA as well as
across SMSA=s. Its economic interpretation, however, is complex. A high
market share by a firm in an SMSA can be an indicator of ability to
differentiate its offerings and, thus, lowers competition (γ1 < 0); on the
other hand, it can also be an indicator of dominance by a firm with a
competitive fringe, which increases competition (γ1 > 0). While our model
says nothing about how the level of competition in a market area is
determined, Ansari, Economides and Ghosh (1994) develop a model in which
either outcome can arise in equilibrium depending on the nature of consumer
20
preferences over attributes. When preferences are nonuniform, the second
outcome is more likely to arise.
Imposing all these functional forms on (8) and (9), we have
[p-{exp(θS)[α1 + α11 βQ(β -1) + α12 D][v1 π(1) v2
π(2) ]} - (σ0 + σ1S4 )] [δ1 ( Q/p )
]
+ Q (1 /(1 + eγM )) = u1
(8)=
[p-{exp(θS)[α1 + α11 βQ(β -1) + α12 D][v1 π(1) v2
π(2) ]} - (σ0 + σ1S4 )] [δ2 (Q/D) ]
+ ( eγM /(1 + eγM ))(r0 + r1 t) - {exp(θS)[α2 + α22 λD(λ -1) + α12 Q][v1 π(1) v2
π(2)
]} = u2 (9)=,
where the only new term is (r0 + r1 t). That is, we have replaced the
shadow price of distribution services, r, by a linear function of the
opportunity cost of time, t , which was measured by the between SMSA
variation in the index of labor compensation.16
V. Results.
Estimation by nonlinear two stage or three stage least squares
requires specifying an instrument matrix. We included every variable
treated as exogenous in the model as an instrument, which gave rise to 16
variables in the instrument matrix, and is similar to what would be done in
the linear case. In the nonlinear case, however, the use of squares of the
original variables and interaction terms improves the efficiency of the
estimates, by making the nonlinear approximations more accurate, although
if carried to an extreme, adding as many variables as one has observations
for example, it can lead to inconsistent estimates.17 We selected six
21
variables that vary both within and across SMSA=s and introduced their
squares as instruments.18 In addition, we introduced a variable not used
earlier and its square as instruments: the percentage of families with two
or more earners in the zip code area where each supermarket was located.
Finally, we added interaction terms between selling area and the other six
variables mentioned in this paragraph and between the percentage of
households with more than two earners and the remaining five variables. To
conclude we used the same 35 instruments, when the squares and interaction
terms are included, in the estimation of the demand equation and the first-
order conditions.19
In Table 3 we present the results of estimating (8)= and (9)= by
nonlinear three stage least squares for each of the waves. Our most
statistically robust results are for the parameter estimates associated
with the two output variables in the cost function. The substantive
implications of these parameter estimates also represent the most important
results of our estimation. The estimates of λ and β suggest that
economies of scale with respect to distribution services and not with
respect to output are the main source of increasing returns to scale for
supermarkets. The null hypothesis that marginal costs with respect to
output (Q) are constant can not be rejected at the 1% level of significance
in any of the three waves, although the point estimate indicates slightly
declining marginal costs. On the other hand the null hypothesis that
marginal costs are constant with respect to distribution services (D) is
rejected at the 1% level of significance and above for all three waves in
22
favor of the alternative that marginal costs are declining. These results
are also consistent with the literature cited in the introduction.20
Finally, a higher level of turnover (Q) increases the marginal costs of
providing a given level of distribution services (D) and viceversa.
>>>>>>Table 3 GOES HERE>>>>>>>>
Our estimates of the shadow price of distribution services are
statistically significant at the 1% level and above for every wave. If for
simplicity we evaluate the estimates at t=0, they imply that the value over
a year to all the consumers of a supermarket of a one unit increase in the
distribution services index varies between $22,100 and 27,800 across the
waves.21 To put this number in perspective note that this represents
between 0.32% and 0.33% of average supermarket sales in the sample, which
were close to 7 million in 1982$. The construction of our index of
distribution services implies that a one unit increase in the index at the
sample mean corresponds to an 11% increase in any of the 20 components
making up the index, i.e., the average value from adding the 20 services ,
which is 11.2357, divided by the average value of the index, which is
101.40. Adding a whole service category implies increasing the
distribution services index by 9, which implies a value to consumers
between 2.88% and 2.97 % of sales.22
With respect to the competition variables, market share and market
growth are statistically significant at the 1% level or above in all three
waves but the number of stores per 1000 persons is not statistically
significant at this level. The coefficient of market growth suggests that
23
the ability to retain patronage by supermarkets is substantially greater in
areas where demand is increasing very rapidly. The coefficient of market
share, on the other hand, suggests that an increase in the market share of
a firm in a market area lowers its ability to retain patronage, perhaps due
to increased competition from the fringe. This effect, however, is much
smaller in magnitude than the effect of market growth.
Of the shift parameters in the cost function, the presence of
scanners generates the most stable results. It is statistically
significant at the 1% level in all three waves and their presence implies a
lowering of costs between 2 and 3 %. The differences in costs between
traditional stores and warehouses are not statistically significant at any
reasonable level, but the differences in costs between warehouses and
superstores are at the 2.5 % level in waves 2 and 3. They imply that
superstores experience between 5 and 6 % higher levels of costs than
warehouses. Of the input parameters, the coefficient of occupancy cost is
positive and statistically significant at the 2.5% level but the
coefficient of labor costs while positive is not statistically significant
at any reasonable level.
Finally, our attempt to compensate for the lack of data on the cost
of goods sold was successful in the following sense. The estimated value
of σ0 is positive and statistically significant at the 2.5% level in waves
1 and 2 and belonging to a chain systematically lowers the costs of
acquiring goods in all three waves. Perhaps more importantly, the bottom
of the table shows that using these estimates there are no violations of
24
second-order conditions at any data points in waves 2 and 3 and only 1
violation in wave 1. The estimated value of the price cost margin index at
the sample mean varies between 23 and 26 index units across the waves.
Expressed relative to the store price index, (p - CQ -w)/p, it also
implies a margin between 23% and 26% of sales.23
We performed two experiments that tested alternative specifications
of competition in the model. Namely, we reestimated the model assuming
first that µ = 0 and subsequently that µ = 1, i.e., the extreme cases of
monopoly and perfectly competitive behavior. In neither case were we able
to obtain convergence, which suggests that imposing these assumptions leads
to a misspecified model. In contrast to these specification tests, we also
performed other specification tests that did not substantially affect any
of the results reported in Tables 2 and 3.24
We also considered an econometric issue especially relevant to our
context. Moulton (1990) argues that cross-section estimates of the effects
of aggregate variables , for example our SMSA variables, on micro units,
the stores in our sample, will be biased if there is spatial correlation.
We took the estimated residuals from (8)= and (9)= and calculated for each
SMSA the test statistic developed by Anselin and Kelejian (1997) for the
case of endogenous regressors with no spatially lagged dependent variables.
This statistic (TS) is defined as
TS = {[v= W v ]/ s2 ]/ tr (WW + W=W=), where v is a an Nx1 vector of
residuals from an equation for an SMSA with N observations; s2 is the
sample variance of the residuals from an equation within an SMSA; W is the
25
spatial weighting matrix, which in our case takes the form of an NxN matrix
with zeros in the diagonal and ones= in the offdiagonal elements. TS has a
χ2 distribution with one degree of freedom. The null hypothesis of zero
spatial correlation could not be rejected at the 1 % level for any of the
28 SMSA=s in either of the two equations.
Finally, we used the estimated model to explore the effects of an
increase in competition, either through a 1 % increase in the price
elasticity of demand (in absolute value terms) or a 1% increase in µ.
We solved the model for p and D for each observation using the values of
our parameter estimates. We augmented the coefficient values as indicated
above and solved the model again for p and D for each observation. This
allowed us to calculate the change in price and the change in distribution
services. We were also able to evaluate these changes in terms of monetary
units. That is, r∆D - ∆pQ (= V) is a money metric utility measure of the
change in consumer welfare. This is especially valuable when changes in p
an D occur in the same direction, because assessing welfare implications
requires an evaluation in terms of monetary units which is provided by V.
Two conclusions emerged from this experiment. First, different
supermarkets respond in different ways to these changes. Some change
prices and distribution services in the same direction, indeed they can
both increase or decrease; some in opposite directions. Second, the four
characteristics (statistically significant at the 5% level) differentiating
the sample of supermarkets where consumer welfare increased (V>0) from
26
those where it decreased (V<0) were a larger turnover (Q), a larger market
share in the SMSA for the firm owning the supermarket, a higher proportion
of superstores and a higher presence of scanners. Hence, these results
suggest that recent trends toward the adoption of superstore formats and
optical scanners are welfare improving for supermarket costumers when
there is an increase in either intratype competition or general competition
for the consumer=s dollar.
Notes
27
* Earlier versions of this paper were presented at the AEA Meetings in Washington, DC,
and at the third conference on Retailing and Services in the Gold Coast, Australia. Seminar
presentations were given at U. of Miami, U. of Maryland, U. of Vienna and the Social Science
Research Center in Berlin (WZB). We would like to thank without incriminating the following
individuals: James MacDonald, Walter Oi, I. Prucha, H. Kelejian, A. Sen, R. Kranton, L. Locay
and N. Wallace. Part of the research for this paper was undertaken while Betancourt was an IRIS
Fellow.
1. E is a restricted expenditure or cost function in that D plays the
role of a fixed input in this function. It has the well known implication (
Shephard=s Lemma) that ME/Mp = q(p, D, p=, z0). This is the Hicksian demand
function which upon substitution of the demand for the commodities, z0 =
g(p, p=, D,Y), generates the Marshallian demand function. See Deaton and
Muellbauer (1980, Chs. 2 and 10) for the general procedure or Betancourt
and Gautschi (1992) for its application to retail demand. In particular
the last reference shows that, if we assume that other prices faced by the
representative consumer (p=) are constant, the Marshallian demand can be
written as q(p, D, Y) and it will be decreasing in price (p), and
increasing in distribution services (D) and the consumer=s full income (Y).
Finally, just as any restricted cost or expenditure function, it has the
property that ME/MD < 0. And, this property can be used to define the
shadow price of distribution services, r = - ME/MD, or how much the
28
consumer would be willing to pay for an additional unit of distribution
services if it were available in the market at an explicit price.
2. If in the above example µ equals zero, we have the case of a
monopolist and there are no gains from meeting the competitive standard.
If µ equals unity the supermarket gains all the business of the
representative consumer. For values between zero and unity the supermarket
gains a fraction of the business of the representative consumer. If µ
exceeds unity second-order conditions are violated. Thus the maximum value
of µ is unity.
3. Incidentally µ captures intratype competition from other supermarkets. General competition
for the consumer=s dollar is captured in the price elasticity of demand. Unless otherwise stated
competition will refer to the former and not the latter.
4. Tirole (1988, p.244) notes that this is the only consistent
conjecture in a static model.
5. For applications of variants of this function to the demand for
supermarket products see Bode (1990).
6. There is a certain amount of judgement in what variables one selects
as X1 and X2 . The first variable suggested itself pretty easily in
light of arguments that the poor pay more at
supermarkets, MacDonald and Nelson (1991). There is no prior literature to
guide the choice of the second one, or to provide expectations about
29
its sign; hence, we chose an indicator of well- being with
independent variation that was not going to be used elsewhere in the
analysis.
7. The latter should be viewed as introduced in exponential form in (4), e.g., as eδX .
8. The list of instruments is the same as used in Section 5 and it will be discussed there.
9. In the equations below the constraint in (3) is not present. Since
the source of the errors is in the application of the decision rules not in
the objective function, the constraint is assumed to hold with certainty.
Note that it is not directly observable or estimable by the
econometrician.
10. The slopes of the demand function in (8) and (9) can be expressed in
terms of elasticities as (MQ/Mp) = δ1 (Q/p) and (MQ/MD) = δ2 ( Q/D), where
δ1 and δ2 correspond to the price elasticity of demand and the distribution
services elasticity of demand estimated in the previous section,
respectively.
11. See the data appendix for additional details.
12. We also considered a different functional form for the cost function,
namely the one employed by Pulley and Braunstein (1984). This function is
the standard quadratic with interaction terms between input prices and
outputs added to allow for heterotheticity. The results deteriorated and
we abandoned experimentation with alternative functional forms.
30
13. In our empirical work the market area was defined as the SMSA in
which the supermarket was located.
14. The variables in the vector M can be thought of as capturing
variations in E=, the minimum expenditure at alternative establishments,
across supermarkets in different market areas.
15. We also considered the SMSA=s four firm concentration ratio and four
firm Herfindhal index. Since these variables, just like the previous two,
take on the same values for each supermarket in an SMSA, they only vary
over the 28 different SMSA=s and this limited variation resulted in severe
multicollinearity problems.
16. Setting r1 to zero made no difference to the substantive results
reported in Section V.
17. See Kelejian and Oates (1981, Ch. 8) for an elementary but insightful
discussion of this issue and Amemiya(1985) for an advanced treatment.
18. Namely the index of labor compensation, the selling area of the
supermarket and four variables available for the zip code area where each
supermarket was located: the percentage of households without cars,
socioeconomic status, median income and population.
19. We checked the sensitivity of our results to the choice of
instruments. For instance, we dropped the interaction terms between the
percentage of households with more than two earners and the six other
31
variables and reestimated the demand equation and the first-order
conditions with the reduced set of 29 instruments. No substantive
statement made about the coefficient estimates in Section III or in this
section would be altered by using these results instead of those based on
the 35 instruments.
20. For instance, Ofer (1973) measured output as value added and argued
that, under certain assumptions, it is a perfect measure of the
(distribution) services provided by a store, D; he found evidence of
increasing returns to scale. On the other hand, Ingene (1984) measured
output as sales per establishment and argued that, under certain
assumptions, it was a perfect measure of turnover or explicit output, Q; he
found evidence of constant returns to scale.
21. Sales were measured in hundred of 1982$. Thus, we multiplied the estimate by 100.
22. Parenthetically, Messinger and Narasinham (1997) have estimated the
average value of one-stop shopping at supermarkets to be between 2.24% and
2.37 % of sales.
23. The average retail margin for food stores in 1982, according to the
Census of Retail Trade, was about 22%. Our numbers, however, are estimates
of the price-cost margin and our cost variables do not account for any
equipment costs.
24. For instance, we changed the values of α1 and α2 from 1 to .5 and
32
1.5 and reestimated the model for all three waves with these new values.
Also, we added the excluded observations from each wave and reestimated the
demand function and the first-order conditions.
References
Amemiya, T.(1984) Advanced Econometrics. Cambridge: Harvard University Press.
Anselin, L. and Kelejian, H.(1997) ATesting for Spatial Error Autocorrelation in the Presence of
Endogenous Regressors.@International Regional Science Review 20, 153-82.
Anderson, K.B. (1990) "A Review of Structure Performance Studies in Grocery Retailing."
Bureau of Economics, Federal Trade Commission. Washington, DC.
Ansari, A., Economides, N. and Ghosh, A. (1994) "Competitive Positioning in Markets with
Nonuniform Preferences." Marketing Science 13, 248-273.
Baumol, W. and Braunstein,Y.(1977) AEmpirical Study of Scale Economies and Production
Complementarities: The Case of Journal Publication.@ Journal of Political Economy 85,
1037-48.
33
., Panzar, J. and Willig, R. (1982) Contestable Markets and the Theory of Industry
Structure. New York: Harcourt, Brace, Jovanovich, Inc..
Betancourt, R. and Gautschi, D. (1992) "The Demand for Retail Products and the Household
Production Model: New Views on Complementarity and Substitutability." Journal of
Economic Behavior and Organization 17, 257-75.
.(1993) "Two Essential Characteristics of Retail Markets and Their Economic
Consequences." Journal of Economic Behavior and Organization 21, 277-94.
Bliss, C.(1988) "A Theory of Retail Pricing." Journal of Industrial Economics 36, 372-91.
Bode, B. (1990) Studies in Retail Pricing. Ph. D. Thesis. Erasmus University.
Cotterill, R.W. (1991)"A Response to the Federal Trade Commission/Anderson Critique of
Structure-Performance Studies in Grocery Retailing." University of Connecticut, Food
Marketing Policy Report No. 13.
Deaton, A. and Muellbauer, J.(1980) Economics and Consumer Behavior. London:Cambridge
University Press.
Divakar, S. and Ratchford, B. (1995) "Estimating the Supply and Demand for Retail Services."
mimeo, SUNY at Buffalo.
34
Ehrlich, I. and Fisher, L.(1982) AThe Derived Demand for Advertising: A Theoretical and
Empirical Investigation.@American Economic Review 72, 366-88.
Ingene, C. (1984) AScale Economies in American Retailing: A Cross-Industry Comparison.@
Journal of Macromarketing 4, 49-63.
Kaufman, P. R. and Handy, C. R.(1989) "Supermarket Prices and Price Differences: City, Firm,
and Store-Level Determinants." Economic Research Service, U.S. Department of
Agriculture, Technical Bulletin No.1776.
Kelejian, H. and Oates, W.(1981) Introduction to Econometrics: Principles and Applications.
New York: Harper & Row Publishers.
MacDonald, J. M. and Nelson, Jr., P. E.(1991) "Do the Poor Still Pay More? Food Price
Variations in Large Metropolitan Areas." Journal of Urban Economics 30, 344-59.
Messinger, P. and Narasimhan, C.(1997) "A Model of Retail Formats Based on Consumers
Economizing on Shopping Time." Marketing Science 16, 1-23.
Moulton, B. R.(1990) AAn Illustration of a Pitfall in Estimating the Effects of Aggregate
Variables on Micro Units.@ Review of Economics and Statistics 72, 334-38.
Ofer, G.(1973) AReturns to Scale in Retail Trade.@ Review of Income and Wealth 19, 363-84.
35
Oi, W. (1992) "Productivity in the Distributive Trades: The Shopper and the Economies of
Massed Reserves." in Z. Griliches (ed.) Output Measurement in the Service Sector.
Chicago, IL: University of Chicago Press.
Progressive Grocer. (1987; 1993, April).
Pulley, L. B. and Braunstein, Y. M. (1984) "Scope and Scale Augmenting Technological Change:
An Application in the Information Sector." in M. Jussawalla and H. Ebenfield (eds.),
Communication and Information Economics: New Perspectives. Amsterdam: Elsevier
Science Publishers B.V.
Tirole, J.(1988) The Theory of Industrial Organization. Cambridge, MA: The MIT Press.
U.S. Bureau of the Census. (1991, 1998) Statistical Abstract of the United States. Washington,
DC.: Department of Commerce.
TABLE 1: DESCRIPTIVE STATISTICS
VARIABLE
MEAN
STD. ERROR
MINIMUM
MAXIMUM
Price wave 1
99.24
5.71
66.36
118.85
Price wave 2
99.66
5.42
73.72
120.92
36
Price wave 3
99.82
5.83
68.76
135.27
Output1 wave 1
704. 47
528.42
52.86
3836.48
Output1 wave 2
702.64
534.47
51.62
3811.87
Output1 wave 3
699.98
528.74
51.66
3772.14
dist. services2
101.40
24.75
8.90
160.20
% wo cars
12.29
13.48
0.
90.
Socec. stat. idx.
54.20
11.23
30.
99.
Median Income
19082.65
6013.86
5901.
51175.
Sh. center dum.
0.6186
0.4863
0.
1.
Population
31586.72
17034.07
189.
93751.
Selling area
18397.20
11234.71
3000.
68381.
Occ. Cost
100.95
23.62
63.70
167.37
Lab. cost3
98.90
23.94
13.54
163.50
Superstore du.
0.1674
0.3738
0.
1.
Traditional du.
0.7883
0.4089
0.
1.
Scanner du.
0.2465
0.4315
0.
1.
Chain dummy
.6140
.4874
0.
1.
Market sh.
0.0801
0.0848
0.0001
0.4661
37
# sts.(per 000's)
0.9792
0.2094
0.6483
1.7838
Market growth
0.0082
0.0425
-0.0900
0.1060
opp. cost4
-0.280
21.28
-37.50
54.66
1 We constructed the output indexes by taking annual store sales in hundred=s of $1982 and dividing by the store
price index.
2 We constructed the distribution services index. Its main difference from the services index constructed by ERS is
the elimination of one of the twenty services categories (scanners, which is an input not an output) and the addition
of another available elsewhere in the original data (extended hours, which is an output); see the Data Appendix for
details.
3 Within SMSA labor cost.
4 Between SMSA labor cost
38
TABLE 2: NONLINEAR TWO STAGE LEAST SQUARES - [DEMAND EQUATIONS]
WAVE 1
WAVE 2
WAVE 3
PARAMETER
ESTIMATE
STD
ERROR1
ESTIMATE
STD
ERROR1
ESTIMATE
STD
ERROR1
δ10 price
-2.1336*
.4496
-2.2276*
.4525
-2.0194*
.4229
δ11 price *
hh wo cars
.0039*
.0006
.0040*
.0006
.0038*
.0006
39
δ20
services
.5016*
.1677
.5221*
.1586
.4639*
.1567
δ21 serv.*
socec. sta.
-.0017*
.0008
-.0019*
.0009
-.0015**
.0008
δ41
shopping
center
-.0433
.0508
-.0450
.0505
-.0403
.0508
δ3 median
income
.9336*
.1829
.9708*
.1733
.8842*
.1705
δ42
population
.0535**
.0316
.0553**
.0315
.0568**
.0316
δ43 selling
area
.8592*
.0504
.8583*
.0483
.8703*
.0481
R2
.515
.536
.525
δ1 MEAN
-2.0856
-2.1785
-1.9727
40
MINIMUM
-2.1335
-2.2276
-2.0194
MAXIMUM
-1.7825
-1.8676
-1.6774
δ2 MEAN
.4094
.4192
.3827
MINIMUM
.3332
.3340
.3154
MAXIMUM
.4505
.4651
.4189
1 White robust standard errors. * |t|>1.96. **�|t|>1.645.
41
TABLE 3: NONLINEAR THREE STAGE LEAST SQUARES -FIRST-ORDER CONDITIONS1
Wave 1
Wave 2
Wave 3
Parameter
Estimate
Std Error2
Estimate
Std Error2
Estimate
Std Error2
θ1 super
-.0385
.0496
.0100
.0364
.0240
.0287
θ2 trad.
.0162
.0465
.0459
.0357
.0584*
.0188
SHIFT
PARAMETERS IN
COST FUNCTION
θ3 scan.
-.0263**
.0152
-0.212*
.0107
-.0178**
.0099
INPUT
π1 occu.
.0720*
.0315
.0606*
.0239
.0524*
.0227
PARAMETERS
π2 lab.
.0271
.0206
.0209
.0188
.0220
.1860
α11
21.16*
8.380
36.76*
13.93
35.81*
14.08
β turno.
.9701*
.0373
.9577*
.0274
.9834*
.0190
42
α22
1633.90
1021.33
1217.76**
722.17
1087.49*
538.372
γ serv.
.4567*
.1359
.5393*
.1393
.5819*
.1251
α12
.0692*
.0153
.0684*
.0133
.0686*
.0131
σ0
35.34*
10.70
25.15**
15.05
17.92
15.72
WHOLESALE
PRICE
PARAMETERS σ1 chain
-1.6044*
.6732
-1.0385*
.5293
-1.5991*
.6010
γ1 sh.
.5140*
.2111
.4716*
.2272
.4182**
.2450
γ2 # st.
-.1154**
.0674
.0141
.0896
.0876
.0790
COMPETITION
PARAMETERS γ3 m.gr.
-1.6642*
.5888
-1.7869*
.6429
-1.6101*
.6170
r0
220.51*
62.95
250.98*
76.40
278.16*
84.27
SHADOW PRICE
PARAMETERS r1 time
.1873**
.0994
.1996*
.0988
.1817**
.0981
Mean
25.98
23.03
25.04
Minimum
-9.73
4.74
5.65
Maximum
52.75
51.24
60.22
p - CQ - w
Obs.<0
1
0
0
1 These estimates were obtained given values of unity for α1 and α2 and using the estimates of δ1 and δ2
for each
wave presented in Table 2.
2 White robust standard errors. * |t| >1.96. ** |t| > 1.645.
43